Mining Uncertain Data Streams Using Clustering Feature Decision Trees

During the last decade, classification from data streams is based on deterministic learning algorithms which learn from precise and complete data. However, a multitude of practical applications only supply approximate measurements. Usually, the estimated errors of the measurements are available. The development of highly efficient algorithms dealing with uncertain examples has emerged as an new direction. In this paper, we build a CFDTu model from data streams having uncertain attribute values. CFDTu applies an uncertain clustering algorithm that scans the data stream only once to obtain the sufficient statistical summaries. The statistics are stored in the Clustering Feature vectors, and are used for incremental decision tree induction. The vectors also serve as classifiers at the leaves to further refine the classification and reinforce any-time property. Experiments show that CFDTu outperforms a purely deterministic method in terms of accuracy and is highly scalable on uncertain data streams.